Analysis, design, and selected applications of multiple winners-take-all networks

As an extension of winner-takes-all to multiple selections, K-Winners take-all (KWTA) is a fundamental operation with widespread applications in sorting, filtering, decoding, clustering, classification, and so on. In this talk, the KWTA problem is formulated as several optimization problems with reducing complexity. Several recurrent neural networks will be presented for solving the formulated problem. In particular, a novel KWTA network with a single state variable and a Heaviside step activation function will be presented. The KWTA network is shown to be globally convergent in finite time. Derived lower and bounds of the convergence time will be discussed. In addition, the initial state estimation will also be delineated for expedition of the process. Extensive simulation results will be delineated and applications to parallel sorting and rank-order filtering will be discussed.